Automatic lexical stress detection using acoustic features for computer-assisted language learning

This paper proposes an English lexical stress detec- tion approach using acoustic features. The approach classifies the vowels of English words into two patterns: primary stress and unstress. We firstly choose the frame-averaged basic feature set of the individual syllable nucleus in polysyllabic words as the baseline to decide the stress pattern. This feature set includes the semitone, the duration, the loudness and the emphasis feature. Furthermore, we introduce the pitch-variation feature set and the context-aware feature set to describe the inside variation characteristic and outside contextual characteristic of the syllable nucleus. By combining the three feature sets, the accuracy rate is improved by 7% 8%. Besides, we train support vector machines (SVMs) classifier for each vowel phoneme respectively. The results show that the phoneme-dependent models performance better than only one shared model. Finally, our system achieved an accuracy of 88:6% compared with human-tagged labels.